Recommendation Systems

Approaches

Collaborative Filtering:

• User-based: Find similar users, recommend their items
• Item-based: Find similar items to what user liked
• Matrix factorization (SVD, ALS)

Content-Based:

• Recommend based on item features
• TF-IDF for text, embeddings for images
• User profile from past interactions

Hybrid:

• Combine collaborative + content-based
• Netflix Prize winner used ensemble

Cold Start Problem

Interview Q: "New user, no history. What do you recommend?" A:

1. Popular items (trending)
2. Ask onboarding questions (genres, preferences)
3. Demographic-based (age, location)
4. Content-based from initial clicks

Interview Q: "Design Netflix recommendations" A:

• Train: Matrix factorization on (user, movie, rating)
• Features: Genre, actors, watch time, time of day
• Ranking: Two-stage (candidate generation → ranking)
• Cold start: Popular in region, onboarding quiz
• Evaluation: CTR, watch time, A/B tests

:::